CN110232334A - A kind of steel construction corrosion recognition methods based on convolutional neural networks - Google Patents

Abstract

The invention discloses a method for recognizing corrosion of steel structure based on convolutional neural network. 3: Perform hyperparameter optimization and model training through cross-validation; Step 4: Input the image of the steel structure to be identified into the model obtained in Step 3 to obtain the corrosion identification result. The invention utilizes the convolutional neural network to realize the automatic extraction of structural corrosion features, avoids complicated and tedious feature design work, improves the efficiency of steel structure corrosion identification, realizes accurate identification of steel structure corrosion, and can provide objective identification results , which provides a new solution for the corrosion identification of steel structures.

Description

A Corrosion Recognition Method of Steel Structure Based on Convolutional Neural Network

technical field

The invention belongs to the technical field of image processing and pattern recognition, and in particular relates to a method for recognizing corrosion of steel structures based on a convolutional neural network.

Background technique

Rust is a frequently occurring defect in steel structures. When the components are corroded, the mechanical properties of the components will be degraded, and the bearing capacity of the structure will also be reduced. Regular rust inspection of the structure will help us understand the health of the structure and take rust removal measures in a timely manner to ensure the safety of the structure.

Human-based visual inspection is currently the most commonly used method for rust detection. Although it is easy to perform, the detection efficiency is low and the detection cost is high. Machine vision-based methods can efficiently perform rust image recognition. However, the recognition accuracy of traditional methods depends on the quality of the designed features. However, high-quality features usually require personnel with professional knowledge to repeatedly conduct feature design experiments to obtain them. Also, the generalizability of artificially designed features is poor.

With the development of deep learning technology, there is a better solution to the problem of steel structure corrosion detection. Deep learning algorithms can automatically extract categorical features from images, thus avoiding complex and tedious feature design work. Moreover, the recognition accuracy based on deep learning algorithms is often higher than that based on traditional methods.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for identifying corrosion of steel structure based on convolutional neural network in order to overcome the deficiencies in the prior art. Compared with the prior art, the invention can significantly improve the corrosion detection efficiency and identification accuracy of the steel structure, reduce the detection cost, and provide objective detection results. At the same time, the method has high applicability and can realize rapid, efficient and accurate corrosion identification of steel structures.

The purpose of this invention is to realize through the following technical solutions:

A method for recognizing corrosion of steel structure based on convolutional neural network, comprising the following steps:

Step 1: Establish a steel structure corrosion picture data set, and divide the training set and the test set;

Step 2: Establish a convolutional neural network model for steel structure corrosion identification;

Step 3: Train the convolutional neural network model: First, use cross-validation to optimize the convolutional neural network hyperparameters; then use the images in the training set to train the convolutional neural network model to obtain the parameters of the convolutional neural network model;

Step 4: Steel structure corrosion identification: Input the pictures in the test set into the convolutional neural network model obtained in step 3, and obtain the recognition result of the convolutional neural network model.

Further, step 1 specifically includes the following steps:

(101) Collect pictures of the steel structure in different time periods and under different lighting conditions through image acquisition equipment;

(102) cutting the picture into a picture of a specified pixel size;

(103) performing data cleaning, removing background pictures and severely blurred pictures; marking the pictures obtained after cutting as rusted or without rusting; performing flipping and scaling operations on the cut pictures to expand the data set;

(104) According to the ratio of 8:2, the data set is divided into training set and test set.

Further, step 2 includes:

Convolutional neural network models can automatically learn features for classification from images and maintain invariance to operations such as translation, scaling, and flipping of images; By adjusting the structure of the existing convolutional neural network model, and setting the number of nodes in the output layer of the existing convolutional neural network model to 2, the structure of the convolutional neural network model for steel structure corrosion identification is obtained.

Further, step 3 includes the following steps:

(301) Optimize convolutional neural network model hyperparameters

Divide the training set obtained in step 1 into five parts on average, use one copy each time as a validation set, and use four copies of the convolutional neural network model established in step 2 to train; repeat five times, using different data each time as validation Set; under the same hyperparameter combination, the mean value of the accuracy of the convolutional neural network model on the validation set is used as the accuracy corresponding to the set of hyperparameters; the hyperparameter corresponding to the highest accuracy is used as the hyperparameter of the convolutional neural network model.

(302) Train convolutional neural network model parameters

Randomly initialize the parameters of the convolutional neural network model, use the images in the training set for forward propagation each time, use the cross entropy function to calculate the loss function generated by the forward propagation, calculate the gradient based on stochastic gradient descent, and use the chain rule for reverse Propagation, update the parameters of the convolutional neural network model; set the training stop condition, and complete the training of the convolutional neural network model. After the training is completed, the training data is input into the convolutional neural network model, and the prediction result of the convolutional neural network model on the training data is obtained, and the training accuracy is obtained by comparing with the real category of the training data.

Further, step 4 includes the following steps:

(401) the pictures in the test set are input into the convolutional neural network model trained in step 3, and the recognition results of the pictures in the test set by the convolutional neural network model are obtained;

(402) Compare with the real category of the pictures in the test set to obtain the prediction accuracy of the convolutional neural network model;

(403) By comparing the training accuracy, the prediction accuracy and the expected accuracy, and the expected accuracy is set manually, it is determined whether it is necessary to add samples or improve the convolutional neural network model to continue training.

Compared with the prior art, the beneficial effects brought by the technical solution of the present invention are:

1. The method for recognizing the degree of corrosion of steel structures based on convolutional neural networks provided by the present invention, firstly uses the corrosion pictures of steel structures to train a deep learning model, and then utilizes the trained convolutional neural network models to perform corrosion identification on the pictures in the test set. During the training process, the convolutional neural network model can automatically learn features for structural corrosion identification from images, thereby avoiding the process of manually designing features and improving the efficiency of feature engineering. The convolutional neural network is not affected by subjective factors in the process of learning features, so it can learn global features for rust classification from the training set, so that the convolutional neural network model has higher recognition accuracy and stronger generalization performance . At the same time, the use of convolutional neural networks can obtain more objective recognition results. The steel structure corrosion identification method based on convolutional neural network has higher accuracy and efficiency than traditional methods, and provides a new solution for structural corrosion identification.

2. The method for recognizing the corrosion degree of a steel structure based on a convolutional neural network provided by the present invention can realize the corrosion recognition of the steel structure with high precision. And according to the detection requirements, the number of nodes in the output layer of the convolutional neural network model can be dynamically adjusted to realize the identification of the corrosion degree of the steel structure. Combining the development with the deep learning framework, the steel structure corrosion identification model can be deployed to the cloud, PC or mobile terminal, so that the steel structure corrosion detection can be carried out at any time, and the detection results can be obtained quickly.

Description of drawings

Fig. 1 is the flow chart of the steel structure corrosion degree identification based on convolutional neural network in the present invention;

2 is a schematic structural diagram of a convolutional neural network model in an embodiment of the present invention;

3 is a flowchart of training a convolutional neural network model in an embodiment of the present invention;

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

As shown in Figure 1, it is the flow chart of the corrosion degree recognition of steel structure based on convolutional neural network of the present invention, including the following steps:

Step 1: Collect steel structure corrosion pictures and establish a steel structure corrosion picture data set, and divide training set and test set;

Step 2: Design the convolutional neural network model architecture according to the data set size, image size and number of image categories;

Step 3: Perform hyperparameter optimization through cross-validation, use the training set data to train the convolutional neural network model, and obtain the parameters of the convolutional neural network model;

Step 4: Input the pictures in the test set into the convolutional neural network model obtained in step 3, and output the recognition results of the pictures in the test set by the convolutional neural network model.

Specifically, step 1 includes:

Step 1.1: Collect pictures at different heights of the steel structure by using image acquisition equipment such as drones, mobile phones or cameras, cut the collected pictures into 224*224 pictures, and remove background pictures and blurred pictures. In order to avoid the loss of image information caused by image segmentation, the number of pixels in the length and width directions of the image is compressed to 224 to an integer multiple. In this embodiment, the picture whose pixel is 4096*3072 is scaled to 4032*2912;

Step 1.2: Mark the image obtained after data cleaning, mark 0 if there is no rust, otherwise mark it as 1. In this embodiment, the training set and the test set are divided according to the ratio of 8:2.

Step 1.3: Deep learning models usually require a large amount of data for training. When the training data is too small, the trained model is prone to overfitting, that is, the classification accuracy on the test set is much lower than the classification accuracy on the training set. Data augmentation on the training set is an effective strategy for mitigating overfitting. In the present invention, the pictures are flipped horizontally, vertically and randomly to expand the data set.

Specifically, step 2 includes:

Convolutional neural networks can automatically learn categorical features from images. The network structure can be designed by changing the position, number, size and number of convolutional layers and pooling layers and other parameters. This example adopts the method of fine-tuning the existing convolutional neural network to obtain a convolutional neural network for steel structure corrosion identification. ZFNet is a commonly used convolutional neural network model. In this example, the number of nodes in the output layer of ZFNet is adjusted to 2, and the other layers are kept unchanged. The structure of the fine-tuned ZFNet is shown in Figure 2.

Specifically, step 3 includes:

The training of the convolutional neural network model consists of two steps: training the hyperparameters of the network and training the network parameters. The network structure used in this example is ZFNet, and the hyperparameters that need to be trained are mainly learning rate, batch training size, random discard rate, and optimizer.

Step 3.1: Train hyperparameters using five-fold five-fold cross-validation method. In each round, the training set is divided into five parts, and one of them is used as the validation set, and the remaining 4 parts are used as the training set. The convolutional neural network is trained using a specified set of hyperparameters, and the convolutional neural network is obtained on the validation set. accuracy. By changing the validation set and training set in turn, the accuracy of 5 different networks on the validation set can be obtained. Repeat 5 rounds, and take the average of 25 precision values as the precision corresponding to the set of hyperparameters.

Step 3.2: The flow of network parameter training is shown in Figure 3. First load the convolutional neural network structure designed in step 2, then randomly initialize the parameters of the convolutional neural network, and set the hyperparameters of the network to the values determined in step 3.1. Read in the training set data, obtain the model's predicted category of the training set data through forward propagation, compare it with the true category of the training set, and use the cross entropy loss function to calculate the error of this forward propagation; use the batch stochastic gradient descent algorithm to The error is back-propagated, and the parameters of the convolutional neural network model are updated layer by layer. When the set training stop condition is reached, the training is stopped and the model parameters are saved.

Specifically, step 4 includes:

Load the test set, network structure, and parameters, and choose the classification performance evaluation criteria, including recall, precision, and F-measure. The recognition results and performance indicators of the convolutional neural network model on the test set can be obtained.

To sum up, the present invention provides a simple method for the corrosion identification of steel structures by using the convolutional neural network model, greatly improves the efficiency of corrosion identification of steel structures, reduces the cost of steel structure detection, and can provide more Accurate and objective test results.

The present invention is not limited to the above-described embodiments and deep learning models used. The above description of the specific embodiments is intended to describe and illustrate the technical solutions of the present invention, and the above-mentioned specific embodiments are only illustrative and not restrictive. Without departing from the spirit of the present invention and the protection scope of the claims, those of ordinary skill in the art can also make many specific transformations under the inspiration of the present invention, which all fall within the protection scope of the present invention.

Claims (5)

1. a kind of steel construction based on convolutional neural networks corrodes recognition methods, which comprises the following steps:

Step 1: establishing steel construction corrosion image data collection, divide training set and test set；

Step 2: establishing the convolutional neural networks model for steel construction corrosion identification；

Step 3: training convolutional neural networks model: convolutional neural networks hyper parameter being optimized using cross validation first； Then using the picture training convolutional neural networks model in training set, the parameter of convolutional neural networks model is obtained；

Step 4: the picture in test set steel construction corrosion identification: being input to the convolutional neural networks model obtained by step 3 In, obtain the recognition result of convolutional neural networks model.

2. a kind of steel construction based on convolutional neural networks corrodes recognition methods according to claim 1, which is characterized in that step Rapid 1 specifically includes the following steps:

(101) picture of the steel construction under different time sections and different illumination conditions is acquired by image capture device；

(102) picture is cut into the picture of specified pixel size；

(103) data cleansing is carried out, background picture and seriously fuzzy picture are rejected；By the picture indicia after cutting be corrosion or Non-corroding；Picture after cutting is overturn, zoom operations, with amplification data collection；

(104) according to the ratio of 8:2, data set is divided into training set and test set.

3. a kind of steel construction based on convolutional neural networks corrodes recognition methods according to claim 1, which is characterized in that step Rapid 2 include:

Convolutional neural networks model can automatically from picture study to the feature for classification, and keep translation to image, The invariance of scaling and turning operation；It is adjusted by the structure to existing convolutional neural networks model, by existing convolution mind Output layer node quantity through network model is set as 2, obtains the convolutional neural networks model knot that identification is corroded for steel construction Structure.

4. a kind of steel construction based on convolutional neural networks corrodes recognition methods according to claim 1, which is characterized in that step Rapid 3 the following steps are included:

(301) optimize convolutional neural networks model hyper parameter

It will obtain that ensemble average is trained to be divided into five parts in step 1, use portion to collect as verifying every time, use four parts of training steps The convolutional neural networks model established in 2；It repeats five times, different data is used to collect as verifying every time；By same super ginseng Under array is closed, the mean value of precision of the convolutional neural networks model on verifying collection is as the corresponding precision of this group of hyper parameter；It will most Hyper parameter of the corresponding hyper parameter as convolutional neural networks model in high precision.

(302) training convolutional neural networks model parameter

Random initializtion is carried out to the parameter of convolutional neural networks model, carries out positive biography using the picture in training set every time It broadcasts, calculates the loss function that forward-propagating generates using entropy function is intersected, gradient is calculated based on stochastic gradient descent, uses chain type Rule carries out backpropagation, updates convolutional neural networks model parameter；Training stop condition is set, convolutional neural networks mould is completed Type training.After the completion of training, training data is input in convolutional neural networks model, obtains convolutional neural networks model to instruction The prediction result for practicing data, is compared by the true classification with training data, obtains training precision.

5. a kind of steel construction based on convolutional neural networks corrodes recognition methods according to claim 1, which is characterized in that step Rapid 4 the following steps are included:

(401) picture in test set is input in the trained convolutional neural networks model of step 3, obtains convolutional Neural net Recognition result of the network model to picture in test set；

(402) it is compared with the true classification of picture in test set, obtains the precision of prediction of convolutional neural networks model；

(403) by comparing training precision, precision of prediction and expected precision, the expected precision is by being artificially arranged, it is determined whether It needs to add sample or improves convolutional neural networks model and continue to train.